I. INTRODUCTION
D (Linz and Donawitz) converter process is executed to raise the bath temperature and reduce the impurity level by blowing oxygen into the steel bath surface and adding appropriate amount of flux and coolant into the bath. The main raw materials of the process include main materials (such as hot metal, scrap, pig iron) and sub-material (oxygen, iron ore, lime, dolomite ), and the product is the steel bath of which the temperature and content are required to hit the tapping aim. The ways of LD converter refining control can be divided into two major groups: Static control based on the information obtained before the start of the blowing and dynamic control based on the process signals representing the state of refining during the blowing. The control method combining the static process control and the dynamic process control is popularly used. Static process control determines the gross requirement of oxygen and coolant for the each heat based on the initial information,
Manuscript received June 26, 2006. This work was supported in part by the URASM CSC Welding and Control center, Algerian Ministry of Scientific Research. H. Meradi is researcher with the Iron and Steel Applied Research Unit Welding and Control Research Center, URASM CSC, BP 196 Annaba, 23000, Algeria (phone: 00 213 38 87 18 81, e-mail: meradi213@yahoo.fr) S. Bouhouche is senior researcher with the Iron and Steel Applied Research Unit Welding and Control Research Center, URASM CSC, BP 196 Annaba, 23000, Algeria (phone: 00 213 38 87 18 81, e-mail: bouhouche11@yahoo.fr). M. Lahreche is the director of the Iron and Steel Applied Research Unit Welding and Control Research Center, URASM CSC, BP 196 Annaba, 23000, Algeria (phone: 00 213 38 87 18 81).
120
values is taking into account for the next heat. The principle of adaptation is shown on Fig. 3.
To define the charge conditions and oxygen blowing requirements to achieve the temperature and composition, mathematical and thermodynamic models have been developed. The LD process is of course a very complex batch reaction, which can vary significantly from heat to heat, while many of variables involved are not accurately known. As a consequence, therefore, it is necessary to have a number of coefficients which are used to correct for differences between the theoretical predictions of the process models and the actual results, in order to improve prediction model. The principle of charge calculation program developed, as show in Figs. 2 and 3, is given for determination of the weight of iron ore, the weight of burnt lime and the volume of oxygen required to meet the weight, temperature and chemistry specifications of a given product. The left side of Fig. 2 shows the information required for the charge calculation: hot metal analysis and temperature; scrap weight, analysis; ore and flux analysis; and weight, temperature and chemistry specification of the product. In the center there is a box labeled Charge calculation program. This program, represents both the computer program and the computer, is often referred to as black box by operators.
III. PROBLEM DEFINITION The developed model, suffer from a degree of imprecision. This is due to inherent complexity of the process, the difficulty of accurate measurement of process variables, and the influence of many factors, such as oxygen lance wear, vessel profile, thermal losses (empty converter) and slag retention, that cannot be quantified. These problems cause the need of reblow. Such reblows add significantly to process times and costs, and results in additional wear of the vessel lining refractory. Alternatively, too high temperature at the end of the blow, leads again to increased refractory wear, poor alloy yield and increase in processing time to reduce temperature to acceptable levels for the subsequent casting process. Approximately 50 % of reblow ratio are due for correction steel bath temperature as shown on the Fig. 4.
60 50,4 50 Reblow Rate (%) 40 30 20 10 0 4,17 3,16 1,66 1,12 0,42 0,23 0,19 36,06
Causes of Reblow
The conventional model developed, use closing for oxygen and heat. After the complete heat, the model recalculates using the actual additions to determine the actual loss (or gain) in oxygen and heat, and therefore the charge to parameters needed to keep the output of the model with current operating conditions. The difference between actual and predicted
With the sight of these results, we decided to develop a more advanced model based on neural networks for temperature control in a first step.
121
C ; C P ;M n; P
C ; C T ;M n; T
n; T
M n
IV. NEURAL NETWORKS MODELLISATION Artificial Neural Networks (ANNs) have become a powerful tool for many complex applications such as function approximation, optimization, non-linear system identification, and pattern recognition. This is due to the fact that they are capable to learn from examples and to perform non-linear mappings. It is confirmed that neural networks can reduce considerably the task of engineers by allowing an effective and generic approach of nonlinear problems [6]. The backpropagation has been applied to a wide variety of practical problems and it has proven very successful in its ability to made nonlinear relationships. A typical backpropagation net is show in Fig. 5. Data are used to develop the model. Before modelling, the data base is divided into three parts, that is, training set, validation set and test set. The training set is used to train a model, the validation set is used to validate the model in order to prevent the model from overfitting, and test set is used to test the model in order to select an accurate model.
V. ARCHITECTURE CONFIGURATION DATA COLLECTION AND PREPROCESSING The preprocessing of data consists in the identification of the relevant information that permits the distinguishing of different classes. Consequently, the emphasis is placed on the elimination of such problems as noise, erroneous or missing data. If too many factors are considered, it is difficult to design a practical model. On the other extreme, the model couldnt reflect the nature of steelmaking if certain important factors are overlooked. Before constructing a model based on industrial process data, its important to indicate how faulty data points might be detected. To locate these faulty data points, four main techniques were used [7][10]: - basic (maximum, minimum and correlation) - structured (analysis of similar input vectors) - multivariate (principal component analysis) - learnt detection (model based analysis) Visualisation and statistical tools are provided to make this task very easy. According to this assumption and after treatment of data base and the choice of most significant variables [11], results are presented in Table II. These variables are the input vector in neural model.
TABLE II INPUT AND OUTPUT RANGES FOR THE DATA SD IS THE STANDARD DEVIATION
Variable Mn cast iron Fig. 5 Typical backpropagation net The neural technique is a powerful method of fitting nonlinear function to data, however, the number of hidden layers and iterations used to train the network should reflect the problem complexity. The difference in vocabulary between statistics and neural networks can summarize in Table I.
TABLE I EQUIVALENCE BETWEEN STATISTICS AND NEURAL NETWORKS
Min. Max. Mean SD Input Input Input 1,76 0,31 3,83 3,04 1,59 5,81
Unit
2,51 0,21 weight % 0,86 0,13 weight % 5,07 0,26 weight % C ton ton Nm3 ton ton
Temperature Input cast iron Weight cast Input iron Weight scrap Input Volume of oxygen Weight lime Input Input
STATISTICS Choice of functions family for approach the function of regression Observations Estimate of parameters for approximation the regression function Interpolation, extrapolation
Weight Input 0 2,5 0,66 limestone Wear of Input 17 918 480 lining refractory Wait Input 4 2160 119 between heats Steel Output 1590 1746 1687 temperature
218 Minute 21 C
Several configurations have been tested, and the best is given by the following diagram:
122
E wij
(4)
wij = x j
(5) (6)
i = f ' (I i ) j wij
j
The simulation was carried out from 1200 heats distributed as follows: 900 heats for training 200 heats for validation 100 heats for test. Algorithm of Training In order to find the best non-linear fit of the output to the data, the following steps must be taken: (i) Choose the architecture of the network, i.e. the network inputs (the relevant variables), the topology and size of the network: this determines a family of non-linear functions with unknown parameters (the weights of the network) which are candidates for performing the data fitting. (ii) Train the network, i.e. compute the set of weight values that minimize the approximation error over the data set used for training (training set). (iii) Assess the performance of the network on a data set (called test set) which is distinct from the training set, but which stems from the same population. The backpropagation is the popular algorithm used in practice [12][17]. This network is called backpropagation because the errors in the network are fed backward, or backpropagated, through the network. This model has been adapted to problems in steel industry in the present work, and also has found lots of applications in various domains of science and technology [18]. The input to any neuron i without bias is given by:
The j components represent the errors of the elements in the following layer, while wij represents the connection weights for these elements. The error of a neuron of the output layer is obtained via:
i = f ' (I i )( yi Oi )
(7)
The error is first of all calculated and then back-propagated into the hidden layer located before the output layer. The connection weights can then be modified according to the calculated wij in the concluding stage of this process. To resolve a problem of local minimum of the error space, we have introduced a momentum term. The equation for adaptation of the weights is modified as follows: (8) wij (t ) = i x j + wij (t 1) is defined as the momentum , t is the current learning step and t-1 the previous learning step. VI. SIMULATION RESULTS After training, the capacity of the network is examined using other data in the so-called test. The network is presented by new examples which also contain output information (as in training). The procedure corresponds to that of the training phase, but the data are not used for learning. In this test, the most important information is the error generated by the network. If this considered to be too large, training should be continued. If the results achieved by the network are satisfactory, it can be used for problem solving. The data used here no longer contain desired output values. They now consist of the input information only. This action on the network is called test. The obtained results from 100 heats are showed in Fig. 7 which illustrate the predicted and measured temperature, Fig. 8 illustrate the error between predicted and measured temperature. Optimal values found for momentum factor and learning rate are 0.3 and 0.7.
I i= wij x j
j
(1)
where Wji are the weights of interconnects between neuron i and j, xj represent the signal at the connection concerned. The sum is transformed by the activation function. The overall transfer function of a neuron is thus structured as follows:
Oi = Ai = f wij x j j
(2)
In this equation Oi is the output of the neuron, Ai is its activation, xj is identical to the output of the preceding neuron, with index j of the observed element. The aim of the learning process is to minimize the global network error:
E=
1 2 ( yi Oi ) 2 i
(3)
123
T(observed) T(predicted)
7 13 19 25 31 37 43 49 55 61 67 73 79 85 91 97 Trial number
VII. CONCLUSION A nonlinear observer is shown to correctly estimate final temperature of molten metal in the Linz and Donawitz (LD) converter process. The mean absolute error was 8 C for the neural network, compared with 30 C for the conventional technique. This therefore represents a very significant improvement. A mixed neural network and knowledge base modeling method shows a very satisfactory result that is much better than the algebraic model. In our approach, the conventional model was used to charge calculation and the neuronal model for temperature prediction in end blow. This results show that the neural network can be successfully used in modelling of LD converter in complement with the conventional model. The objective of the first stage is achieved; we project to generalize this method for process control (prediction bath temperature and composition of steel in end blow). REFERENCES
[1] [2] [3] [4] G. Poferl et al., Dynamisation of the LD process (first results), second EOSC 1997. H J Whittaker et al., The use of neural networks in BOS process modelling, ECSC workshop Brussels, 22 and 23 January 1998, pp. 13544. Xianwen Gao et al., Intelligent compound control of direct current electric arc furnace, IFAC 1997, pp. 135-40. Shujiang Li et al., Neural network techniques and its application in ladle furnace burden, IFAC, korea, 1997, pp. 165-68. J. Poferl, Dynamic refining LD converter, ISIJ, Vol 28, 1988, pp. 59-67. M. V. Spangler, (1994). Uses of Process Insights at a Refractory Gold Plant, Pavilion Users Conference, October 17, 1994, pp. 1-7. G. Dreyfus, les rseaux de neurones, Mcanique Industrielle et Matriaux, n51 (septembre 1998). S. J. Qin and al.: Advances in instrumentation and control; Proc. Conf, Austin, TX, USA, October 1993, ISA, vol. 48, N3, pp. 1711-1720. J. Tenner and al.: Intelligent processing and manufacturing of materials; Proc. Conf. IPMM 99; Honolulu, july 1999. J. Tenner; Prediction of mechanical properties in steel heat treatment process using neural networks; Ironmaking and steelmaking, 2001, Vol. 25, pp. 15-22. H. Gutte; the mathematical simulation of the LD process; technical report, TU-BAF, Germany, January 2000. J.-P. Corriou; Les rseaux de neurones pour la modlisation et la conduite des procds, Lavoisier, 1997. K. Andersen et al; IEEE Trans. Ind. Appl., 26 (1990) n5, pp. 824-30. Bhatt et al ; IEEE Contr. Syst. Mag., 10 (1990), pp. 24-29. P. Bhagat; Chem. Eng. Progr., 86 (1990), pp. 55-60. W. E. Staib et al; Iron steel Eng., 69 (1992), pp. 29-32. M. A. Reuter et al; Metall. Trans., 23 B (1992), pp. 643-650. A. Datta and al.; Adaptive neural net models for desulphurization of hot metal and steel; Steel research , 65 n11, (1994), pp. 466-471. Data Engine; Overview and User Manuel; MIT, 2nd Edition, May 1999.
Error
Temp erature (C )
Trial number
Fig. 8 Error between predicted and measured temperature
[5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19]
A mean absolute error of 8 C with a standard deviation of 7 C is achieved. A further important consideration is the performance of the neural technique in comparison with the conventional prediction model. Taking end blow bath temperature prediction, Fig. 9 shows the performance of prediction bath temperature between conventional and neural net model. The mean absolute error was 8 C for the neural network, compared with 30 C for the conventional technique. This result represents a very significant improvement.
1800 Predicted temperature 1750 1700 1650 1600 1 3 5 7 9 11 Heat number 13 15
Fig. 9 Comparison conventional model versus neural net prediction of bath temperature
124